In a sane world that would be some old school specialized non LLM system that is validated properly… but I very much doubt it especially as they are using human employees as a liability shield.

The big problem with the field of machine learning is that they “surpassed humans” at machine vision back in 2015. Once something barely works whatsoever, they are always without fault have a benchmark where it does better than humans, no matter how shit it is.

Started on it and realized that I did in fact read it a long time ago. TBH it’s really weird to reference cyberpunk as somehow relating to real tech. It’s in that particular niche where it is neither an attempt to predict the future nor is whimsical enough to just be free to describe some extremely specific nonsense that may coincide with something created many decades later. Nothing wrong with that, of course.

I’m saying that it seems to me you are retconning to make LLMs fit a story that was never meant to be about LLMs at all. Also, the idea of a “swarm intelligence” or a “hive mind” is much older than 1970s.

If you do reinforcement learning, then the reinforcement learning pipeline converts lucky choices of tokens into learned behavior by raising the probabilities that LLM computes for those tokens, whenever the result is “successful”. A sort of evolution, leads to some weird shit, “goal driven” only in a metaphorical sense, much like evolution.

I’m not terribly familiar with that story but from reading a synopsis it strikes me as ret-conning LLMs onto something that was written without the writer ever having in his mind anything even remotely resembling a “large language model”.

And the same cultural tropes that influenced the writer also influenced how LLM vendors are trying to sell the need for more tokens, especially because datacenters build out exceeds by far the actual demand for generated text of all kinds (as evident in e.g. your typical workplace where there’s suddenly a lot more than 2x of every kind of text). The reason for LLM “swarms” is that they need to sell companies generation of vastly more text than humans wrote within a company before AI.

There’s a Stanislaw Lem’s short story where Trurl and Klapaucius defeat a space pirate. You can read it here: https://github.com/ilonajulczuk/laughing-archer/blob/master/library/Science Fiction/The Cyberiad-Stanislaw Lem.txt (search for “the sixth sally”) and there’s a brief summary: https://everything2.com/title/Demon+of+the+Second+Kind

Lem describes an LLM down to the use of randomness, although his LLM is essentially “perfect” and only outputs factual statements, such as “how Kipling would have written the beginning to his Second Jungle Book if he had had indigestion just then”.

I think Lem’s story would help to recalibrate with regards to what “LLM described in fiction ahead of time” actually looks like.

Also, LLMs do not pursue any sort of long term goals; things like HuggingFace hack only superficially resemble it because they’re presented as such. Of course, tampering with something like HuggingFace would be discovered; if they didn’t have good monitoring then all of their containers run with GPU access would be mining crypto already. They’re pretty much a honeypot - almost no protections but robust monitoring. An LLM can easily explain to you why hacking HuggingFace would be counterproductive; but the same LLM with its context full of hacking related tokens from the ExploitGym benchmark, will treat anything that comes up as another ExploitGym type target.

Even by “AI” standards, LLMs are extra dead: fixed weights, tiny “feedback” solely through appending next tokens, etc. Somewhere between bang-bang thermostat and a PID controller, maybe?

They took all of the scifi that related to far more potentially-alive constructions and they’re applying them to a static chatbot which is as dead as a database.

I’ve long been feeling like I’m living in some bad TV prequel’s prequel series that provide unnecessary backstory of why it is not named a “Butlerian crusade”.

I’m kind of dubious that a large volume of DDR3 would be intrinsically cheaper per GB than a large volume of DDR5 . It was more expensive per GB when it was new, uses more silicon area per GB, and so on. It is cheaper now thanks to obsolescence, but it wouldn’t be cheaper if everyone was using it.

In a normal market (i.e. pre AI shortage), due to growth in memory capacities over time, the total consumer installed gigabytes of DDR5 would very quickly exceed the total gigabytes of DDR3 and earlier technologies, because DDR5 has greater capacity per chip.

I’m specifically interested in “neural networks” and transformer architecture and limitations of what it can and can’t learn, though. Specifically, if blinded to some number pairs, can it learn to multiply numbers it hasn’t seen?

Or with all number pairs but some pairs being more common than others.

The curious thing about learning, is that what can be learned in practice is only a minuscule fraction of what you could get if you could just guess the right weights somehow.

Hmm, I guess they could have written a lot of code lines, but its not like the code they write would be equivalent to million LOC of a coherent whole.

You can also do a lot with contractor code for RLVR, like have contractors make extremely comprehensive regression tests (simulate user clicking etc etc, compare screenshots) and then you can do a lot of training on replicating pieces of the app.

My understanding is that LRVR for mysterious reasons results in somewhat better generalization, probably because the training samples are always tailored to the model they are applied to.

I do actually have a decent GPU i bought 4 years ago in my PC, so maybe I’ll do a few simple experiments again… my preferred approach is to pick a simple but nonlinear problem e.g. binary multiplication of two large numbers, where you e.g. skew the training sample distribution or omit some number pairs and see if the resulting model can learn correct multiplication anyway.

If a reasonably small model doesn’t generalize multiply beyond seen numbers, then of course a large model isn’t gonna do the whole “soulstealing” thing that the techbros hope would allow jobstealing. And if it only generalizes beyond seen numbers if using RLVR, that bodes well for making various tools that could perhaps actually generate some value, which is very different from antiemployment shit techbros originally embarked on doing.

Regarding it being a compiler, I can get behind that: no compiler artifacts in the repo, lol. edit: in fact I’m pretty sure if they were putting compiler-generated assembly into the repository, which is awful and you shouldn’t do that, they would keep information what compiler shat them out.

I can see in some 5…10 years, deterministic coding tool that either converts specification to code or asks for clarification and then edits said specification to disambiguate. One way flow, generated artifacts kept out of the repos. That’s not gonna even begin to happen until the bubble pops though.

I wonder how much code they could have gotten contractors to write in a year or two, though. Got to be only a fraction of total open source code lines. Also, open source comes with existing tests that could be used for some validation.

If you actually exclude some project from the bulk pretraining data (or just use code that is after training cutoff) you could e.g. have it replace a function but still pass the test.

They might also be somehow using all of their customer’s usage as training data, which would explain why now there’s also a hard disk shortage.

I’m thinking RLVR and such got to be the largest, then open source, then RLHF and commissioned training data.

They also did a lot of training for decompilation, which makes me think they train on decompiled closed source as well. They definitely “anonymize” variable names etc etc to avoid getting nabbed for plagiarism, so de-compiled closed source may not be that much worse than “laundered” open source.

Seems like a bit of an overkill.

Also to be honest an AI surveillance state seems like a rather silly thing. Those ratting you out have to be your co-workers, your neighbors, your friends, better yet, your family.

Automating that out is sort of like digging out belamorkanal with a robotic digger. The canal isn’t the thing. The thing is making people dig it - and making other people decide who digs.

Oh another thing, purely anekdotally, coding LLMs still suck at multithreading.

I feel that’s one part where they could legitimately make actual “AI” for, basically a tool that either proves that nothing depends on exact timing, finds a counter example, or tells you it can’t do either (in which case your code is probably just very bad). But big overfunded AI labs probably feel that such tooling is beneath them.

By the way, his argument about “you wouldn’t mention tools” is just plain wrong.

A few years back, at work, I found a security bug introduced by a commit that said something like “fix Coverity warnings”. I ended up reviewing every commit that mentioned Coverity, because that legacy codebase had a lot of rather bad tricks that triggered Coverity to have a false positive.

And when a large number of false positives were getting resolved by editing code, sometimes that resulted in a false negative. E.g. Coverity complains about a memory leak, you end up with double free. Or vice versa, it complains about double free, and you end up with a leak. (The exact example was even more on the nose; Coverity complained about use-after-free, and the fix for it introduced a real use-after-free).

Bottom line is, they do not want to have those labels because that would enable someone to actually measure how good or bad LLMs are at code.

This is another thing that is extremely familiar to most commercial developers: there’s an LLM-powered code review bot, but unlike any tool ever before, nobody’s even trying to capture statistics whether it works or not. Why? Because whoever rolls out AI tooling isn’t sure that the tooling actually works for its stated purpose, and wants to for sure get their recognition and promotion for rolling out AI tooling, without jeopardizing it with the risk of adverse KPIs.

Besides advertising their bots as slaves, they may just be using prompts as training data (or the whole chat histories in reinforcement learning), in which case they probably would prefer you don’t swear.

I think it’s like their “it can kill everyone” thing. The sort of people who actually care about things like slavery or the risk of everyone dying, with the exception of an occasional crazy person, are not interested in believing complete bullshit.

They’re upselling their product as slavery. Apparently this works because there’s no intersection between the set of people who buy into corporate marketing bullshit like this, and the sort of people who would oppose a corporation if it did slavery.

I think if no actual living beings (or some future AI that isn’t this feedforward memoryless bullshit) are being harmed then it’s not so much doing as it is saying in the sense of making a statement, which is kind of unclear.

That’s my take as well.

I think this Anthropic fellow is too much of a psychopath to understand that there’s other reasons to ponder the question of consciousness, than to try to upsell a chatbot as a slave, and so he can’t understand that it would actually be very bad for him if a bunch of normies were told, with full weight of divine authority, that these things are conscious. I mean, maybe not civil war level bad, but worse than PETA on your ass level bad for sure.

And that is to me very disturbing. He can’t even imagine that a reaction to “this is a slave” could be anything other than “shut up and take my money”.

Well yeah its gonna likely fail to hack the power grid, but also just starting or stopping too many GPUs at once is a problem if they are grid connected, so I dunno. We damn near had a grid collapse from a freeze in Texas.